C
Journal of Environmental Monitoring
Dynamic Article Links <
Cite this: J. Environ. Monit., 2012, 14, 589
PAPER
www.rsc.org/jem
Assessing temporal representativeness of water quality monitoring data Saku Anttila,*a Mirva Ketola,b Kirsi Vakkilainenb and Timo Kairesalob Received 21st September 2011, Accepted 21st November 2011 DOI: 10.1039/c2em10768f The effectiveness of different monitoring methods in detecting temporal changes in water quality depends on the achievable sampling intervals, and how these relate to the extent of temporal variation. However, water quality sampling frequencies are rarely adjusted to the actual variation of the monitoring area. Manual sampling, for example, is often limited by the level of funding and not by the optimal timing to take samples. Restrictions in monitoring methods therefore often determine their ability to estimate the true mean and variance values for a certain time period or season. Consequently, we estimated how different sampling intervals determine the mean and standard deviation in a specific monitoring area by using high frequency data from in situ automated monitoring stations. Raw fluorescence measurements of chlorophyll a for three automated monitoring stations were calibrated by using phycocyanin fluorescence measurements and chlorophyll a analyzed from manual water samples in a laboratory. A moving block bootstrap simulation was then used to estimate the standard errors of the mean and standard deviations for different sample sizes. Our results showed that in a temperate, meso-eutrophic lake, relatively high errors in seasonal statistics can be expected from monthly sampling. Moreover, weekly sampling yielded relatively small accuracy benefits compared to a fortnightly sampling. The presented method for temporal representation analysis can be used as a tool in sampling design by adjusting the sampling interval to suit the actual temporal variation in the monitoring area, in addition to being used for estimating the usefulness of previously collected data.
Introduction Temporal variation can cause a significant source of uncertainty in the water quality monitoring of water bodies.1,2 Seasonal patterns and periodicity that are mainly driven by light, temperature and nutrient availability induce the main features in temporal variation in temperate and polar lake ecosystems.3 However, several factors can cause variability in these cycles. These include external drivers, such as perturbations by human activities and sudden meteorological events and how these a Finnish Environment Institute, Mechelininkatu 34a, PL140 00251 Helsinki, Finland. E-mail:
[email protected]; Fax: +358 9 5490 2690; Tel: +358 400 148732 b University of Helsinki, Department of Environmental Science, Niemenkatu 73, 15140 Lahti, Finland
interact with drainage basin, as well as endogenous processes, such as water currents and movements and interactions of organisms within and between different trophic levels.4–6 It is argued that due to this partly unpredictable temporal variation monitoring programmes, guided by the European Water Framework Directive, collect too few data to produce scientifically justified information for the intended purpose.7,8 In considering the temporal representativeness of monitoring datasets, the problem is obviously associated with the sampling interval and how it is related to the actual temporal variation in the sampled area. Sampling regimes for different water quality monitoring methods vary. For example, the Finnish national monitoring programmes collect samples from predefined locations that range from four times a year to only once in every twelfth year, depending
Environmental impact To rationalize water quality monitoring programmes, sampling intervals need to be calibrated against the temporal variation of a specific monitoring area. High frequency data from automated monitoring stations can be used in this adjustment but usage of such data requires careful calibration and quality control. In this study, we present a method for temporal representation analysis that can be used as a tool in sampling design by adjusting the sampling interval to suit the actual temporal variation in the monitoring area. It can be also used for estimating the usefulness of previously collected data. In general, representation analysis of water quality monitoring data is one step forward in the more controlled usage of water quality monitoring data. This journal is ª The Royal Society of Chemistry 2012
J. Environ. Monit., 2012, 14, 589–595 | 589
on the significance of the respective water body.9 An obvious major limitation is the availability of funding.10 When the number of samples is limited, adjustment to temporal variation is usually done by timing the sampling to a certain seasonal event, such as late winter stagnation, summer stratification or autumnal circulation.11 Although many local surveillance programmes reach fortnightly or even weekly sampling frequency levels, expert judgment will play an important role in the interpretation of such data until temporal variance is better understood.1 High frequency measurements accessible to automated monitoring stations, or buoys, have undeniable advantages. Automated measurements have low costs per measurement and such data can be used to detect sudden events in water quality, and can even be used to provide early warning signals of ecosystem regime shifts.12,13 Automated water quality measurements are also suggested as an approach to overcome the challenges in adjusting sampling intervals to temporal variation.14 Probably the most commonly used method to monitor continuously the amount of phytoplankton is the in vivo fluorescence of photosynthetic pigments, especially chlorophyll a (hereafter referred to as chl-a) in phytoplankton cells. This method was developed as early as the 1960’s.15 Nevertheless there are several factors, which adversely affect the accuracy of automated measurements. A typical problem is that the biomass of cyanobacteria cannot be evaluated by using chl-a in vivo fluorescence alone, since cyanobacterial chl-a is mostly located in the nonfluorescing photosystem I.16,17 Therefore, measuring fluorescence at wavelengths typical for chl-a underestimates the true chl-a concentration in time periods when cyanobacteria are abundant. This problem can partly be corrected by measuring the in vivo fluorescence of phycocyanin (PC), i.e. phycobilin pigments that are specific for cyanobacteria.17 Furthermore, fluorescence from chl-a and PC pigments varies as they are affected by factors such as the species composition, cell size, cell physiology or by specific responses such as to excitation type and spectrum, nutrient availability, light intensity, and other environmental conditions.17,18 Therefore, fluorescence data need careful calibration before they can be used as a measure of chl-a concentration. This study presents a statistical method to characterize the expected accuracy of seasonal mean and standard deviation estimates of chlorophyll-a concentration for different sampling intervals. Furthermore, the probability of different sampling intervals to miss the time periods with high concentrations was analysed. The analysis is based on high frequency automated monitoring data from three stations which had been carefully calibrated and recalculated to match the actual seasonal variation in the monitoring area. The results are compared with typical sampling frequencies of manual monitoring.
Methods Study site and datasets Lake Vesij€arvi is located in Southern Finland (25 370 2400 E 61 00 3000 N). Enonselk€a is the southernmost basin of the lake on which the sampling stations were located. The basin has a surface area of 26 km2 and a mean depth of 6.8 m. The lake is dimictic. Typically the stratification period lasts from June to August, and during the winter (November–April), the lake is covered with ice. 590 | J. Environ. Monit., 2012, 14, 589–595
Fig. 1 Locations of three automated monitoring stations on Enonselk€a basin in Lake Vesij€arvi with general land use classes from the nearby areas.
Lake Vesij€ arvi has a long history of eutrophication, remediation, and monitoring activities.19,20 Currently, the lake is mesotrophic with 20–30 mg P l1 of total phosphorus.20 We used hourly fluorescence data of chl-a obtained from three automated monitoring stations installed on the Enonselk€ a basin of Lake Vesij€ arvi (Fig. 1) and complemented these data with intensive manual sampling conducted next to the stations during May–October 2010. One hourly fluorescence measurement was based on the average of 2000 measurements during 30 second period. At each station the sensors (TriOS Micro Flu chl) measured chl-a fluorescence (chl-a fl) at the excitation and emission wavelengths of 470 and 685 nm, respectively. The Ruoriniemi station also measured the fluorescence of phycocyanin (PC-fl) at the excitation and emission wavelengths of 620 and 655 nm (TriOS Micro Flu blue). Data flow from TriOS Micro Flu chl-sensors was set to transform the fluorescence into chl-a concentrations by using a standard conversion coefficient provided by the manufacturer. A similar standard coefficient for phycocyanin fluorescence was provided by the supplier of the instrumentation (Luode Consulting Ltd. Address: Olarinluoma 15 B, FI02200, Finland). Calibration of automated measurements In 2010, we carried out intensive fortnightly chl-a sampling at the three monitoring stations in order to characterize the accuracy of This journal is ª The Royal Society of Chemistry 2012
chl-a fluorometer measurements and to develop a calibration methodology. Altogether 33 chl-a samples were taken, 11 samples from each of the three stations. Sampling started on 12th of May, and continued fortnightly until 1st of September, after which samples were taken at one month interval until 27th of October. Samples were taken using a Limnos tube sampler at a 1 m depth, next to each fluorometer. In the laboratory, the samples were filtered through Whatman GF/C filters (Whatman International, Maidstone, UK), extracted with ethanol and measured spectrophotometrically. In order to derive the maximum correspondence for chl-a concentrations measured by fluorescence, regression models were developed for each station separately, using laboratory derived chl-a concentrations as a response variable and the fluorescence of chl-a, PC, or both from the automated stations as predictor variables. Only one of the stations (Ruoriniemi) included a phycocyanin fluorometer. Since PC fluorescence is concluded to improve the chl-a in vivo measurements,17 these data were applied as a predictor variable also in regression models for other stations. All the stations are located in the same basin and within 1.2 km distance from each other. To reduce the effect of possible spatial variation during the day, the PC fluorescence data were included in the model as a daily mean. Using both chl-a and PC fluorescence as predictor variables yielded the most accurate validation for chl-a concentration. Therefore, we used the coefficients derived from the multiple linear regression to calibrate the hourly chl-a fluorescence data. The performance of calibration models was validated using an independent dataset obtained from the same sites in 2009. This was achieved by comparing the calibrated fluorescence values to chl-a monitoring data collected next to the stations by the environmental authorities and the University of Helsinki. Finally, daily mean values for the calibrated, hourly measurements for the May to October periods for the years 2009 and 2010 were calculated. These datasets were then used in the temporal representativeness analysis. Temporal representativeness analysis Moving block bootstrapping analysis was applied to each time series in order to derive the standard errors (SE) for the mean and variance estimates of different sample sizes. Bootstrapping methods allow the estimation of the sample distribution of almost any statistic by using a simple method based on a computerized calculation.21 One benefit of this method is that it does not require normality in the dataset, which could hinder the purely statistical estimations of the standard errors of measured environmental datasets.22,23 Moving block bootstrap is an iterative process by which each iteration round time series is divided into equal length blocks according to the sample size variant. Then a random sample is taken from each block and the mean and variance are calculated from the resulting set of samples. Furthermore, it was marked if the set samples did not include observations in the upper 75%-percentile of the original dataset. We iterated a similar random sampling from our daily mean time series 1000 times for each sample size (n) ranging from n ¼ 1 to n ¼ n/2. By using daily mean values, we disregarded the effect of diurnal variation which is obviously a factor in representative sampling analysis. However, in daily mean time series, the highly variable diurnal variation is averaged off and consequently these This journal is ª The Royal Society of Chemistry 2012
time series were considered to give a more adequate picture of the seasonal succession. Finally the standard errors, i.e. the sample standard deviation divided by the square root of the sample size (s/On) for the mean and the standard deviation estimates, were calculated from each set of sample size statistics. Results are presented as relative standard errors in which the derived standard errors of the mean were divided by the mean of the whole time series. Similarly, the standard errors of the standard deviation estimates were divided by the standard deviation of the whole dataset. Also the percentages of iteration rounds without observations in the upper 75%-percentile were calculated for each sample size. The moving block bootstrap calculation was programmed to run under Matlab-software (Matworks Inc.). Relative standard errors of the mean and standard deviations and the percentages without observations in the upper 75%percentile from all time series were then combined into three respective datasets, following which a simple function (eqn (1)) was fitted to the resulting standard errors of different sample sizes: a0 f SEnp ¼ (1) np where np is the percentage sampled from the time series and a0 is the coefficient derived by the nonlinear fitting. Dataset including the percentages of observations without upper quartile observations in different sample sizes was modelled with the exponential function (eqn (2)): 1 f EUQnp ¼ bnpb (2) 0 where b0 and b1 were derived by the nonlinear fitting. These functions were empirically found to describe the combined datasets adequately. Fitting was performed by nonlinear regression in which the SEs were used as the respondents and the Table 1 Coefficients of determination (r2) and root mean squared error (rmse) for simple linear regression models using only chl-a fluorescence (chl-a-fl), PC-fluorescence from Ruoriniemi applied to all stations (PC-fl) and multiple linear regression using chl-a fl from all stations and PC-fl from Ruoriniemi (chl-a and PC-fl). Data were from year 2010 and N ¼ 11 for each model. Interceptions, coefficients (coefs), their standard errors (SE of coef.) and p-values for the regression models are also shown Model
Reg. stats.
Ruoriniemi
Lankiluoto
Myllysaari
chl-a fl
r2 rmse Intercept Coefs b1 SE b1 P-VAL b1 r2 rmse Intercept Coefs b1 SE b1 P-VAL b1 r2 rmse Intercept Coefs b1 Coefs b2 SE b1 SE b2 P-VAL b1 P-VAL b2
0.000 3.642 10.50 0.027 0.269 0.921 0.454 2.691 7.264 1.586 0.579 0.023 0.580 2.370 3.506 0.3 2.003 0.215 0.627 0.2 0.013
0.025 3.929 10.69 0.212 0.381 0.591 0.422 3.026 7.281 1.670 0.652 0.031 0.736 2.045 1.43 0.776 2.518 0.233 0.515 0.010 0.001
0.024 3.458 10.21 0.189 0.337 0.588 0.474 2.539 7.028 1.557 0.547 0.019 0.836 1.416 0.576 0.728 2.361 0.164 0.359 0.002 0.0002
PC-fl
chl-a and PC-fl
J. Environ. Monit., 2012, 14, 589–595 | 591
Fig. 2 Comparison between calibrated automated chl-a measurements from 3 stations (Calib) and laboratory measured chl-a from 2009 (n ¼ 12) with linear regression (lreg) and its coefficient of determination (r2) and root mean squared error (rmse). Calibration was based on the multiple linear regression models derived from the data of 2010.
percentages sampled from the whole time series were used as the predictors. The calculation method, which was run on Matlabsoftware, utilized the Levenberg–Marquardt algorithm in an iterative minimization process to find the least squares residuals between model and observations.
Results Calibration of automated measurements Using only chl-a fluorescence as the predictor variable in estimating the chl-a concentration in 2010 led to poor correspondence
(Table 1). Instead the PC fluorescence from Ruoriniemi explained more of the variability in chl-a concentrations than chl-a fluorescence for the whole season in all of the stations. The best correspondence was obtained when both chl-a fluorescence from each stations and PC fluorescence from the Ruoriniemi station were used as predictor variables. Coefficients obtained from these multiple regressions were used to calibrate hourly chl-a fluorescence data of 2010. When these multiple regression models for each station for the data obtained in 2010 were applied to the automated monitoring data of 2009, the correspondence against laboratory determined chl-a concentrations increased significantly (Fig. 2), whereas the prediction power of chl-a fluorescence alone was basically zero (r2 ¼ 0.03 and rmse ¼ 7.18). Thus, the multiple linear regression models using chl-a fluorescence from each station and PC-fluorescence from the Ruoriniemi station were also used to calibrate the chl-a fluorescence data of 2009. Using phycocyanin fluorescence data in the calibration models clearly improved the accuracy, especially in late summer when phycocyanin fluorescence started to increase, which was indicative of increased concentrations of cyanobacteria (Fig. 3). During early season, the need for the inclusion of PC in the calibration was lower.
Temporal representativeness analysis The standard errors (SE) of the mean and standard deviation against the percentage sampled from the time series are presented in Fig. 4. As expected, an increasing sample size reduced the SE of the mean more rapidly than did the SE of the standard deviation. Fitted models described the simulated SEs of the mean and standard deviation reasonably accurately with the squared correlation coefficient values (r2) of 0.88 and 0.89, respectively. The expected accuracy limits of mean and standard deviation estimates of different sampling intervals are presented in Fig. 5. With monthly sampling the relative standard errors of the mean
Fig. 3 Automated PC-fluorescence measurements from Ruoriniemi station (A) in 2009 and 2010. Chl-a measurements from three stations, Ruoriniemi (B), Lankiluoto (C) and Myllysaari (D) in 2009 and 2010. Measurements calibrated with multiple linear regression (solid line) and uncalibrated chla fluorescence (dashed line) values are plotted together with water sample chl-a concentration (open circles).
592 | J. Environ. Monit., 2012, 14, 589–595
This journal is ª The Royal Society of Chemistry 2012
Fig. 4 Estimated and modeled relative standard error (SE) percentages for the mean (A) and standard deviations (B).
and standard deviations can be estimated to be close to 10% and 23%, respectively. The expected accuracy increases rapidly with increasing sampling frequency. With fortnightly sampling the corresponding errors were approximately 4.5% and 11%, and with weekly sampling only 2% and 5%. The probability of not observing concentrations in the upper 75%-percentile of different sample sizes is presented in Fig. 6. According to the results, even monthly sampling interval (3.3% sampled from the time series) gives a relatively high probability for observing high concentrations (12.8%). Fitted model described the estimated probabilities accurately (r2 ¼ 0.98).
Discussion
Fig. 5 Modeled accuracy limits ( standard error percentages) for the mean and standard deviation estimates of different sample sizes from the time series. Vertical lines indicate the monthly, fortnightly and weekly sampling intervals.
Fig. 6 Estimated and modeled percentages for the probability (prob.) of not observing concentrations (conc.) in the upper 75%-percentile of different sample sizes.
This journal is ª The Royal Society of Chemistry 2012
Most of the national or international water quality monitoring programmes tend to unify the sampling interval in order to standardize the guidelines for monitoring in different areas.7,9,24 The disadvantage of this approach is that it lacks information on whether the collected data can give a representative picture on the water quality in a specific monitoring area. Carefully calibrated high frequency monitoring data from automated measurement stations reveal a temporal variation in water quality. Using such data we presented a method to describe the temporal accuracy of different sampling frequencies for a specific monitoring area. The results showed that in Lake Vesij€ arvi a relatively high error in seasonal statistics estimations can be expected with monthly sampling. On the other hand, weekly sampling did not significantly improve the accuracy when compared to fortnightly sampling. Doubling the sampling frequency can roughly be estimated to double the monitoring expenses as well. Also the probability for missing observations from concentrations in the upper 75%-percentile was relatively small even with low sampling frequencies. These periods with high concentrations typically lasted more than one week in our time series and this explains the high likelihood to get observations from the upper quartile of the time series. The importance of tailoring the sampling frequency to the actual temporal variation has been stressed in the literature. It has been concluded that many water quality monitoring programmes are not sufficient for their intended purpose since the J. Environ. Monit., 2012, 14, 589–595 | 593
sampling frequency regimes they follow are not calibrated to the actual temporal variation.8 A common flaw is that the number of observations used to describe water quality over certain time period is insufficient, which causes a significant uncertainty to inter alia annual mean estimates.1 For instance, in a study in Missourian lake reservoirs it was found that almost a quarter of chl-a observations varied more than 150% from the long term mean.25 Although the problem of insufficient sampling frequencies and the consequent risk of falsely estimating true levels have been clearly identified, we did not find any studies in the literature that present a methodology on how the temporal variation could be taken into account in water quality monitoring. It is clear that meteorological and anthropogenic perturbations cause inter-annual variation and therefore our two year datasets were not sufficient to fully describe typical inter-annual variation of our study site. Nonetheless even less comprehensive datasets can give valuable information for the planning and quotation of monitoring programmes. The presented method can be assumed to be less accurate in water sources that often face sudden and short term changes in water quality. In rivers, for example, meteorological events in drainage basin can drive such rapid changes in water quality that usage of only statistical sampling design is probably insufficient. The advantage of using the moving block bootstrap method as presented in this study is that it is based on measured variation in water quality. Moreover, the block bootstrap method does not require statistical assumptions such as normality for the respective datasets, which can be a problem in the analyses of ecological time series.26 Modeling the error in water quality statistics estimates has many advantages. Measured confidence intervals of existing data are valuable in accuracy assessments of ecological modeling or assimilation of different data sources.27,28 Most importantly, the method used in this study is one step forward in the more controlled usage of water quality monitoring data. It has been stated that if the systematic variation can be more adequately described, then the random variation will be decreased, indicator precision improved and monitoring requirements reduced.1 Temporal representation analysis can be used to rationalize sampling efforts. It can also be used to minimize the risk of oversampling. Oversampling is rarely a problem in the fundinglimited manual sampling programmes, but it might be an issue of concern with other water quality monitoring methods. Remote sensing applications, for example, can reach even daily measurement frequency. In other words, temporal representation analysis can be used as a tool in the sampling regime design to calibrate sampling frequency to the temporal dynamics of the monitoring area in addition to estimation of the usability of collected data. While automated monitoring is increasingly growing and its capabilities are more widely recognized the accuracy assessment of collected data is trailing behind.13 Our in situ technology for phytoplankton quantification, when compared with conventionally used laboratory measurements of chl-a concentration, revealed the importance of calibration by discrete sample collection and also by the contribution of PC-fluorescence in the calibration. In Lake Vesij€arvi, the usage of only manufacturer’s calibration in automated chl-a fluorescence measurements would have given inconsistent errors during the measurement season. In 594 | J. Environ. Monit., 2012, 14, 589–595
early summer, manufacturers calibration gave estimates twice as high as the actual phytoplankton abundance, whereas during autumn the estimates were too low. The increasing biomass of cyanobacteria, as shown by phycocyanin fluorescence, probably caused these changes, as chl-a in cyanobacteria is mostly located in the non-fluorescing photosystem I.17 There are several sources of variation in chl-fluorescence that are not explicitly resolved. For example, lower fluorescence during autumn may be caused by increasing amounts of inactive chl-a in dying cells or cells being shaded by others, as these reduce fluorescence.29 On the other hand, phytoplankton cells adapted to low irradiance such as that which occurs during autumn can have a higher chla content per cell than those adapted to high irradiance. However, it has been concluded that with falling irradiance the total amount of chl-a in cyanobacterial cells decreases and that of phycocyanin increases.3 Despite these confounding factors, measuring the fluorescence of both chl-a and phycocyanin in combination improved the calibration of our in situ fluorescence data. It is preferable that several pigments which are activated at varying light-spectrum wavelengths are needed in order to get as reliable an estimate as possible of the actual phytoplankton biomass.18,30 In general, online measurements do not exclude conventional sampling and these two methods should be considered as complementary.17,18,29 The importance of the careful calibration of automated fluorescence measurements cannot be over-emphasized. To conclude, our results showed that the calibration of fluorescence data using values obtained from discrete water samples is essential in automated phytoplankton monitoring. The usage of PC-fluorescence in the calibration in combination with chlfluorescence was required, at least in our study lake. By applying calibrated high frequency monitoring data, we numerically described the effect of sampling intervals on the accuracy of seasonal statistics estimations. This kind of accuracy assessment can be important in planning and calculating quotations for monitoring programmes, in addition to being a well-founded starting point for the assimilation of different water quality monitoring data sources.
References 1 J. Carstensen, Statistical principles for ecological status classification of water framework directive monitoring data, Mar. Pollut. Bull., 2011, 55, 3–15. 2 T. P. Burt, N. J. K. Howden, F. Worrallc and M. J. Whelan, Longterm monitoring of river water nitrate: how much data do we need?, J. Environ. Monit., 2010, 12, 71–79. 3 R. G. Wetzel, Limnology: Lake and River Ecosystems, Academic Press, NY, USA, 3rd edn, 2001. 4 D. G. George and S. I. Heaney, Factors influencing the spatial distribution of phytoplankton in a small productive lake, J. Ecol., 1978, 66, 133–155. 5 J. Horppila, H. Peltonen, T. Malinen, E. Luokkanen and T. Kairesalo, Top-down or bottom-up effects by fish: issues of concern in biomanipulation of lakes, Restor. Ecol., 1998, 6, 20–28. 6 L. Tuvenike, T. Noges and P. Noges, Why do phytoplankton species composition and ‘‘tradiotional’’ water quality parameters indicate different ecological status of a large shallow lake?, Hydrobiologia, 2011, 660, 3–15. 7 European Communities, Common Implementation Strategy for the Water Framework Directive (2000/60/EC), Guidance document no. 7. Monitoring under the Water Framework Directive, 2003, http:// forum.europa.eu.int.
This journal is ª The Royal Society of Chemistry 2012
8 L. H akansson, A data reduction exercise to detect threshold samples for regression models to predict key water variables, Int. Rev. Hydrobiol., 2007, 92/1, 84–97. 9 S. Mitikka and K.-M. Vuori., Monitoring of Inland Waters, in Environmental Monitoring in Finland, ed. J. Niemi, The Finnish Environment, 2009, vol. 12, pp. 31–32. 10 G. M. Lovett, D. A. Burns, C. T. Driscoll, J. C. Jenkins, M. J. Mitchell, L. Rustad, J. B. Shanley, G. E. Likens and R. Haeuber, Who needs environmental monitoring?, Front. Ecol. Environ., 2007, 5, 253–260. 11 The Finnish Eurowaternet, with Information about Finnish Water Resources and Monitoring Strategies, ed. J. Niemi, P. Heinonen, S. Mitikka, H. Vuoristo, O.-P. Pietil€ainen, M. Puupponen and E. R€ onk€ a, The Finnish Environment, 2001, vol. 445. 12 T. Huttula, E. Bilaletdin, P. H€arm€a, K. Kallio, J. Linjama, K. Lehtinen, H. Luotonen, O. Malve, B. Vehvil€ainen and L. Villa, Ymp€ arist€ on seurannan menetelmien kehitt€ aminen. Automatisointi ja muut uudet mahdollisuudet, Suomen ymp€arist€ okeskuksen raportteja, 2009, vol. 13. 13 S. R. Carpenter, J. J. Cole, M. L. Pace, R. Batt, W. A. Brock, T. Cline, J. Coloso, J. R. Hodgson, J. F. Kitchell, D. A. Seekell, L. Smith and B. Weidel, Early warnings of regime shifts: a whole-ecosystem experiment, Science, 2011, 332, 1079–1082. 14 T. Suominen, H. Tolvanen and R. Kalliola, Geographical Persistence of Surface-Layer Water Properties in the Archipelago Sea, FENNIA, SW Finland, 2010, vol. 188/2, pp. 179–196. 15 C. J. Lorenzen, A method for the continuous measurement of in vivo chlorophyll concentration, Deep-Sea Res., 1966, 13, 223–227. € 16 D. Campbell, V. Hurry, A. K. Clarke, P. Gustafsson and G. Oquist, Chlorophyll fluorescence analysis of cyanobacterial photosynthesis and acclimation, Microbiol. Mol. Biol. Rev., 1998, 62, 667–683. 17 J. Sepp€ al€ a, P. Yl€ ostalo, S. Kaitala, S. H€allfors, M. Raateoja and P. Maunula, Ship-of-opportunity based phycocyanin fluorescence monitoring of the filamentous cyanobacteria bloom dynamics in the Baltic sea, Estuarine, Coastal Shelf Sci., 2009, 73, 489–500. 18 T. L. Richardson, E. Lawrenza, J. L. Pinckney, R. C. Guajardo, E. A. Walker, H. W. Paerl and H. L. MacIntyre, Spectral fluorometric characterization of phytoplankton community compositionusing the Algae Online Analyser, Water Res., 2010, 44, 2461–2472.
This journal is ª The Royal Society of Chemistry 2012
19 T. Kairesalo, S. Laine, E. Luokkanen, T. Malinen and J. Keto, Direct and indirect mechanisms behind successful biomanipulation, Hydrobiologia, 1999, 395/396, 99–106. 20 T. Kairesalo and K. Vakkilainen, Lake Vesij€arvi and the city of Lahti (southern Finland): comprehensive interactions between the lake and the coupled human community, SILnews, 2004, 41, 1–5, http://www. limnology.org/news/silnews41.pdf. 21 B. Efron and R. J. Tibshirani, An Introduction to the Bootstrap, CRC Press, Boca Raton, FL, 1994. 22 R. M. Vogel and A. L. Shallcross, The moving block bootstrap versus parametric time series model, Water Resour. Res., 1996, 32/6, 1875– 1882. 23 H. Varian, Bootstrap tutorial, Mathematica J., 2005, 9, 768– 775. 24 U.S. Environmental Protection Agency, Elements of a State Water Monitoring and Assessment Program, EPA 841-B-03-003, 2003, http://www.epa.gov/owow/monitoring/repguid.html. 25 M. F. Knowlton and J. R. Jones, Temporal variation and assessment of trophic state indicators in Missouri reservoirs: implication for lake monitoring and management, Lake Reservoir Manage., 2006, 22/3, 261–271. 26 A. D. Jassby and T. M. Powel, Detecting changes in ecological time series, Ecology, 1990, 71/6, 2044–2052. 27 S. Marsili-Libelli, S. Guerrizio and N. Checchi, Confidence regions of estimated parameters for ecological systems, Ecol. Modell., 2003, 165, 127–146. 28 J. Pulliainen, J. Veps€al€ainen, S. Kaitala, M. Hallikainen, K. Kallio, V. Fleming and P. Maunula, Regional water quality mapping through the assimilation of spaceborne remote sensing data to shipbased transect observations, J. Geophys. Res., 2004, 109, C12009, DOI: 10.1029/2003JC002167. 29 K. Izydorczyk, C. Carpentier, J. Mr owczynski, A. Wagenvoort, T. Jurczak and M. Tarcznska, Establishment of an alert level framework for cyanobacteria in drinking water resources by using the Algae Online Analyser for monitoring cyanobacterial chlorophyll a, Water Res., 2009, 43, 989–996. 30 J. Gregor, R. Geris, B. Marsalek, J. Hetes and P. Marvan, In situ quantification of phytoplankton in reservoirs using a submersible spectrofluorometer, Hydrobiologia, 2005, 548, 141–151.
J. Environ. Monit., 2012, 14, 589–595 | 595
